{"id":"c9ca2197-96c4-4b4d-9064-8b0853df5656","arxiv_id":"2506.07417","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"EviSEC combines evidential uncertainty and spectrum-aware contrastive training to detect out-of-distribution dynamic graphs, reporting strong AUROC gains over prior baselines.","lead":"EviSEC detects unusual, out-of-distribution samples in evolving networks by converting the model's predictions into uncertainty scores and fabricating unusual training samples from the graph spectrum. On six real-world dynamic graph datasets it reports much better detection than eleven existing baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 14 contrastive loss appears to drive negative samples toward, not away from, the ID prediction, inverting the stated uncertainty-widening mechanism.","rationale":"The reader's weakest assumption is the spectral augmentation heuristic: that removing low-frequency Laplacian components yields OOD approximations whose uncertainty behavior matches real OOD dynamic graphs. That is a plausible empirical generalization concern, but it is about transfer across OOD types and can only be probed with additional experiments. In contrast, the problem I identify in Eq. 14 is an internal inconsistency: the loss, as written, appears to minimize exactly the quantity the paper claims it maximizes. Minimizing Lcl aligns the negative sample's predicted class distribution with the ID sample's sharp distribution, so it should make negative samples more confident, not more uncertain. This is directly checkable from the equations and does not depend on assumptions about spectrum-frequency semantics. The evidential losses L_ce-edl and L_kl are standard, and the claim that argmax ID performance is preserved follows from Eq. 7-8 because the exponentiation and normalization preserve the order of logits. The remaining pillar of the method is the contrastive loss, so if Eq. 14 is wrong in direction, the stated mechanism for enlarging the score gap collapses. The empirical gains in Tab. 2-4 may still be real if the released code optimizes a different objective, but the paper's formal description would then be inaccurate. This is precisely the kind of issue that warrants a conditional verdict: the authors must correct Eq. 14 or verify the implementation, and the code should be committed with a hash so the training objective is reproducible. I therefore retain CONDITIONAL, but for a different and more immediate reason than the reader's spectral-augmentation concern.","tokens_in":13798,"tokens_out":7627,"duration_ms":89259,"concrete_test":"Pin the GitHub repository to a commit and extract the exact implementation of Lcl. Compute the gradient of Lcl with respect to the negative sample's concentration α^- for a fixed positive sample with one-hot p. If the gradient increases α^-_sum (thereby lowering u^-), the implemented loss narrows the ID/OOD gap, contradicting the paper's claim. As a complementary check, re-run the Tab. 4 ablation with Lcl replaced by a loss that explicitly maximizes the entropy of p^- (or equivalently minimizes Σ_i p_i log p_i^- with a minus sign); if performance remains at the reported level, Eq. 14 as written is not the source of the gains, and the paper should be corrected.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central mechanism for mitigating score homogenization is the contrastive loss Lcl in Eq. 14. As written, Lcl = E_{p^-}[Σ_i p_i log p_i^-] is the cross-entropy (up to sign) between the positive sample's class distribution p and the negative sample's distribution p^-. Minimizing Lcl therefore drives p^- toward p. For a representative ID input, p is sharp (near one-hot). The loss then rewards a negative sample that is also sharp on the same class (e.g., for p=[1,0], Lcl = log p_1^-, minimized when p_1^-→1). This makes the negative sample more confident, not flatter; uncertainty u^- = K/Σα^- decreases, narrowing the ID/OOD gap. This directly contradicts Sec. 4.3, which states the loss lets the model 'assign higher uncertainty scores for OOD approximations' and 'widens the gap in uncertainty scores.' If Eq. 14 is exactly what was optimized, the reported gains cannot be attributed to the stated mechanism. If the implementation uses a different loss (e.g., maximizing the entropy of p^- or minimizing p_i^- for the true class), then the paper's formulation is wrong and the experiments need to be re-described. The GitHub link lacks a commit hash and reproduction configuration, so the actual training objective cannot be verified from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EviSEC, an out-of-distribution (OOD) detector for dynamic graphs that combines evidential deep learning with spectrum-aware contrastive augmentation. A dynamic GCN-GRU encoder produces evidence vectors, which are converted into Dirichlet posterior concentrations; the resulting uncertainty u = K / alpha_sum is used as the OOD score. To combat score homogenization, the method removes low-frequency components of the normalized graph Laplacian to synthesize negative samples, and a contrastive loss is added to the evidential losses. Experiments on six real-world datasets with two OOD types (structure manipulation and feature interpolation) report AUROC, AUPR, and FPR95 improvements over eleven baselines, together with ablations and ID F1 checks.","tokens_in":14060,"tokens_out":11293,"duration_ms":133240,"significance":"The application of evidential uncertainty to OOD detection in dynamic graphs is timely, and the six-dataset evaluation with eleven baselines is substantial. The paper provides a clear problem formulation, uses standard evidential losses, and includes an ablation study that separates the contributions of the loss terms. If the central mechanism were correctly specified, the work would be a useful step for dynamic-graph OOD detection. However, as written, the contrastive objective in Sec. 4.3 does not logically enforce the claimed high-uncertainty behavior for negative samples, and the spectral augmentation pipeline is underspecified, so the reported gains are not yet attributable to the stated design.","major_comments":[{"comment":"The contrastive loss is written as L_cl = E_{p^-}[sum_i p_i log p_i^-], which is a log-likelihood (equivalently, negative cross-entropy). Since Eq. (15) is minimized, minimizing L_cl maximizes the cross-entropy between the ID class distribution p and the negative-sample distribution p^-. For an approximately one-hot p at class c, this is equivalent to minimizing E[log p_c^-], which drives p_c^- toward zero; that can be achieved by a sharp posterior concentrated on another class, which gives low uncertainty u = K / alpha_sum, not the flat, high-uncertainty posterior claimed in the same section. If the implementation instead maximizes L_cl, then p^- is pulled toward p and the uncertainty gap narrows. The manuscript must state the optimization direction and, more importantly, either change the objective to one that directly enforces high uncertainty (e.g., a KL-to-uniform or entropy penalty on p^-) or provide an analysis showing that the stationary solution of Eq. (14) is flat. As it stands, the core score-homogenization mechanism is not established.","section":"Sec. 4.3, Eq. (14)"},{"comment":"The augmentation step defines a matrix L_r^- as a sum of outer products of Laplacian eigenvectors, but the dynamic graph encoder in Eq. (5) consumes an adjacency matrix A_t and a feature matrix X_t. The paper does not state how L_r^- is converted into an input graph snapshot: whether the modified Laplacian is used directly in the message-passing step, or whether an adjacency matrix is reconstructed from it. This is needed both for reproducibility and for the claim that the generated matrices are 'OOD approximations' of dynamic graphs. Please specify the full augmentation pipeline, including how node-count variations and multiple temporal snapshots are handled.","section":"Sec. 4.3, Eq. (13)"},{"comment":"The paper reports that grid search determines the three key hyper-parameters (the balancing factors rho_1 and rho_2 and the preservation ratio r) but never gives the optimal values actually used for the results in Tables 2 and 3. The GitHub link also has no commit hash or configuration file. Without these values, the central quantitative claims cannot be reproduced or checked against the sensitivity analysis in Fig. 5. Please report the chosen hyper-parameters per dataset and provide a versioned code release.","section":"Sec. 5.1 / Sec. 5.5"},{"comment":"The spectrum-aware augmentation is justified by the heuristic that removing low-frequency Laplacian components produces OOD-like graphs, but the test OOD sets are generated by stochastic block models and feature interpolation (Sec. 5.1), not by spectral perturbation. The paper offers no analysis (spectral, embedding-level, or otherwise) showing that the augmented negatives resemble the test OOD distribution. If the augmentation heuristic is mismatched to real OOD types, the contrastive training could teach the model a signal that does not transfer. Please add a quantitative comparison of the augmented negatives and the test OOD samples, or explicitly discuss this limitation.","section":"Sec. 4.3 / Sec. 5.2"}],"minor_comments":[{"comment":"The abstract uses 'EviSAC' once while the rest of the paper uses 'EviSEC'; please unify the name.","section":"Abstract"},{"comment":"The section header reads 'OOD Detection Perfrmance'; it should be 'Performance'.","section":"Sec. 5.2"},{"comment":"The ablation table is difficult to parse: the caption '✓-means the variant with random negative edge sampling augmentation' is unclear, and the row '✓ ✓ ✓ -' appears to have an extra column. Please reformat the table so each ablation variant and its configuration are unambiguous.","section":"Table 4"},{"comment":"The reported 'average AUROC improvement of 24.32%' is not defined; please state explicitly that the average is taken over all baselines and all datasets, and clarify whether the percentage points are absolute improvements.","section":"Sec. 5.2"},{"comment":"Figure 4 shows ID F1 scores for only four of the six datasets, while the text refers to it after a general statement about ID performance; please clarify why UCI and AS are omitted or add them.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is reasonably extensive, but the theoretical mechanism in Sec. 4.3 needs correction or careful qualification before the paper can be accepted. As noted in the major comments, the stress-test note's exact inversion claim is not correct, but the underlying issue—that Eq. (14) does not enforce flatness or high uncertainty—remains valid. Please also verify that the released code corresponds exactly to the reported experiments, including the actual sign of the contrastive loss."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe paper applies evidential deep learning (EDL) to OOD detection in dynamic graphs, with a spectrum-aware augmentation to generate synthetic negatives, and reports large AUROC gains over 11 baselines across six datasets. The framing is sensible and the experiments are extensive. But Eq. 14, the contrastive loss at the heart of the method, does not do what the paper says.\n\nWhat's new: combining EDL with a dynamic graph encoder and spectral augmentation for OOD detection is a reasonable integration. The EDL derivations in Sec. 4.2 are standard and correct, and the ablation study shows all three loss components contribute. The ID performance preservation claim is backed by F1 comparisons. So there is a real contribution here, if the mechanism works.\n\nThe problem: L_cl in Eq. 14 is E_{p^-}[Σ p_i log p_i^-]. Minimizing this drives p^- to put high probability on classes where p_i is low—it makes the negative sample confident on a different class, not flat or uncertain. That directly contradicts Sec. 4.3's claim that the loss makes the model 'assign higher uncertainty scores for OOD approximations.' The stated mechanism for mitigating score homogenization is therefore invalid as written. If the implementation optimizes exactly Eq. 14, the reported gains cannot be attributed to the stated mechanism. If the implementation uses a different objective (e.g., entropy maximization), the paper misdescribes it. The code link has no commit hash or reproduction config, so this is not verifiable.\n\nOther soft spots: no error bars or significance tests, which matters given the ~24% average improvement claim; the spectral augmentation justification rests on a single citation; and the eigen-decomposition cost on the largest dataset (Elliptic, 203K nodes) is not discussed.\n\nProportionally, this is a fixable flaw, not a hopeless paper. I would send it to peer review with a strong request to correct or clarify the loss, verify the code, and add uncertainty quantification. But I would not cite it or build on it until that's addressed.\n\nBest,\n[Name]","headline":"Eq. 14's contrastive loss pushes negative samples toward confident wrong classes, not flat ones, contradicting the paper's central uncertainty-widening claim; otherwise a sensible EDL-for-dynamic-graphs OOD paper with extensive experiments.","tokens_in":14572,"tokens_out":7020,"would_cite":false,"duration_ms":83678,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EviSEC claims that out-of-distribution (OOD) detection in dynamic graphs can be recast as an uncertainty-estimation problem, with a Dirichlet posterior's uncertainty serving as the OOD score.","keywords":["dynamic graph","out-of-distribution detection","evidential deep learning","graph spectrum","contrastive learning","Dirichlet uncertainty"],"falsifier":"Run EviSEC with OOD types it never trained on, such as temporal shifts where test snapshots come from a later time period or label distribution shift without structural manipulation, and measure whether the AUROC gain over the strongest baseline persists; if the uncertainty gap learned from low-frequency deletions does not separate these real OOD samples, the core hypothesis is falsified.","tokens_in":13569,"feed_emoji":"🕸️","tokens_out":3552,"duration_ms":35845,"temperature":0.7,"pith_summary":"EviSEC claims that out-of-distribution (OOD) detection in dynamic graphs can be recast as an uncertainty-estimation problem. Instead of relying on a single point prediction, the model outputs a Dirichlet posterior over class probabilities, and the uncertainty of that posterior becomes the OOD score. Because training sees no OOD data, the paper adds a spectrum-aware augmentation that deletes low-frequency Laplacian components of graph snapshots to synthesize OOD-like negatives, and a contrastive loss widens the uncertainty gap between ID snapshots and these negatives. On six real-world datasets with two OOD types, the method reports an average AUROC improvement of 24.32% over eleven baselines while keeping in-distribution F1 close to state-of-the-art performance. The reason to care is that safe deployment of dynamic graph models in security-sensitive settings hinges on flagging inputs that deviate from the training distribution, and this is a parameter-light way to learn such a flag.","feed_headline":"Widening the uncertainty gap catches OOD dynamic graphs","feed_subtitle":"Training with spectrum-cut fake OOD snapshots lifts AUROC by 24.32% on average while keeping in-distribution performance.","key_machinery":"The load-bearing object is the evidential head: exponentiated logits minus one form a Dirichlet concentration vector $\\alpha$, from which class probabilities are the mean of the posterior and uncertainty is $u = K/\\alpha_{\\text{sum}}$. Two evidential losses, cross-entropy on the posterior and a KL penalty toward a uniform Dirichlet, keep ID distributions sharp. The second mechanism is spectrum-aware augmentation: eigen-decomposing the normalized Laplacian, keeping all high-frequency eigenspaces, keeping only the first $r$ fraction of low-frequency eigenspaces, and reconstructing a graph $L^-_r$ that approximates an OOD sample; a contrastive loss then maximizes the separation between the uncertainty of original snapshots and augmentations. Together they convert the missing-OOD-data problem into a self-supervised widening of the uncertainty gap.","core_discovery":"The paper's central claim is that a Dirichlet-evidence output head, trained jointly with a spectrum-based contrastive loss, produces an uncertainty score that separates in-distribution from out-of-distribution dynamic graph snapshots far more sharply than post-hoc scores such as softmax confidence, energy, or plain entropy. The model preserves the argmax class probability during evidential training, so ID classification accuracy is not sacrificed; OOD detection is driven by the total Dirichlet concentration $\\alpha_{\\text{sum}}$, with uncertainty $u = K/\\alpha_{\\text{sum}}$ as the score. The spectral augmentation removes a controlled fraction of low-frequency graph Laplacian eigenspaces to manufacture negative samples, and the contrastive term pushes these negatives toward higher uncertainty. The reported result is an average AUROC gain of 24.32% over baselines and large FPR95 reductions, with near-perfect detection on two bitcoin networks.","pith_inferences":["A natural test the paper leaves implicit: whether the same low-frequency deletion works when the downstream task is link prediction rather than node or edge classification, since contrastive negatives are generated per snapshot regardless of task.","The method suggests a general recipe: any augmentation that produces high-uncertainty synthetic points can replace OOD data, so spectral deletion is one instance of a broader uncertainty-anchored negative mining principle.","One could test whether using the preservation ratio $r$ as a continuous curriculum, starting with mild deletions and increasing severity, improves calibration of the uncertainty score.","Because the paper's OOD generation follows the static-graph framework of GNNSAFE, a direct comparison against real-world OOD examples, such as suddenly appearing fraud rings, would clarify how much of the gain is due to the augmentation matching the test generator."],"forward_implications":["If the claim holds, post-training OOD detectors on dynamic graphs can be replaced by training-time evidential heads plus spectral negatives, yielding large gains on AUROC and FPR95.","The Dirichlet uncertainty score is a single scalar threshold $\\gamma$, so deployment can simply compare $u$ against a learned cutoff.","Because ID F1 is preserved, the same model serves both prediction and OOD flagging without a separate detector.","The spectral augmentation recipe transfers to any dynamic graph encoder that outputs node embeddings, since it operates on the Laplacian of each snapshot."],"supporting_citations":[{"why":"Supplies the evidential deep learning framework, including the Dirichlet posterior and the $L_{ce-edl}$ and $L_{kl}$ loss forms.","marker":"[33]"},{"why":"Provides the graph-spectrum rationale that low-frequency components capture global structure, justifying the low-frequency deletion augmentation.","marker":"[34]"},{"why":"Supplies the EvolveGCN dynamic graph encoder that EviSEC uses as its backbone for temporal snapshots.","marker":"[19]"},{"why":"Defines the SM and FI OOD generation framework used to build test OOD samples and provides the GNNSAFE energy-based baseline.","marker":"[9]"},{"why":"Serves as the strongest uncertainty-based baseline (DAEDL) that EviSEC compares against and outperforms.","marker":"[29]"},{"why":"Supplies the energy-based OOD score baseline used in the comparison experiments.","marker":"[25]"},{"why":"Supplies the maximum softmax probability baseline that motivates the single-point-estimation critique.","marker":"[23]"}],"fun_headline_variants":["Evidential spectrum contrast widens OOD gap","Dirichlet uncertainty catches OOD dynamic graphs","Spectral negatives boost OOD AUROC by 24%","Contrastive evidential training for graph OOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a dynamic graph made OOD-like by deleting low-frequency Laplacian components has uncertainty behavior similar to genuinely out-of-distribution dynamic graphs, so the contrastive signal transfers to unseen OOD types.","fun_headline_variants_meta":{"raw":{"variants":["Evidential spectrum contrast widens OOD gap","Dirichlet uncertainty catches OOD dynamic graphs","Spectral negatives boost OOD AUROC by 24%","Contrastive evidential training for graph OOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1295,"prompt_tokens":981,"completion_tokens":314,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":252}},"tokens_in":597,"tokens_out":314,"duration_ms":3808,"temperature":1.0,"reasoning_tokens":252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:35:23.966170+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EviSEC with OOD types it never trained on, such as temporal shifts where test snapshots come from a later time period or label distribution shift without structural manipulation, and measure whether the AUROC gain over the strongest baseline persists; if the uncertainty gap learned from low-frequency deletions does not separate these real OOD samples, the core hypothesis is falsified.","supporting_citations":[{"cited_title":"Advances in neural information processing systems31(2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the evidential deep learning framework, including the Dirichlet posterior and the $L_{ce-edl}$ and $L_{kl}$ loss forms."},{"cited_title":"Advances in Neural Information Process- ing Systems35, 2972–2983 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the graph-spectrum rationale that low-frequency components capture global structure, justifying the low-frequency deletion augmentation."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelli- gence","cited_arxiv_id":null,"evidence_quote":"Supplies the EvolveGCN dynamic graph encoder that EviSEC uses as its backbone for temporal snapshots."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Serves as the strongest uncertainty-based baseline (DAEDL) that EviSEC compares against and outperforms."}],"review_version":1}